macOS arm64 release binaries require macOS 26 (Tahoe) minimum - breaks on macOS 15 and earlier

[EnzoPython3]

Summary

The macOS arm64 release binaries have LC_BUILD_VERSION set to minos 26.0, meaning they refuse to run on any macOS version before 26 (Tahoe). This silently breaks atomic-agent's managed local backend for anyone not yet on macOS 26.

Environment

• macOS 15.7.8 (Sequoia), build 24G824
• Apple Silicon (arm64), M3 Pro
• atomic-agent, managed mode, model gemma-4-12b

Steps to reproduce

1. Install atomic-agent, leave localModels.mode as "managed"
2. Let atomic-agent download the macOS arm64 backend release into <stateDir>/models/backend/
3. Start a chat turn

Actual behaviour

atomic-agent surfaces only:

Turn failed [transport]: fetch failed

There's no indication that the backend process never launched, or why. It looks like a network/config problem, not an OS compatibility one.

Root cause

$ otool -l llama-server | grep -A4 LC_BUILD_VERSION
     cmd LC_BUILD_VERSION
 platform 1
   minos 26.0
     sdk 26.5

dyld refuses to load a binary whose minimum OS requirement exceeds the running OS, so the managed llama-server process never actually starts, and every subsequent fetch to the local port fails.

This reproduces on both the current release at the time of filing (b10269-1.5.1) and the prior one (b10018-1.3.0, 27 Jul), so it looks like the macOS build job moved to a macOS 26 runner without pinning an explicit deployment target (e.g. -mmacosx-version-min / MACOSX_DEPLOYMENT_TARGET) below the build host's OS version.

Suggested fix

• Pin the macOS build job's deployment target (e.g. macOS 14 or 15) so releases stay runnable on non-bleeding-edge installs, or
• If macOS 26 is an intentional minimum going forward, surface that clearly in release notes and ideally have atomic-agent detect the spawn failure and report the real cause instead of a generic transport error.

Workaround

Installed llama.cpp via Homebrew (bottled for Sequoia), ran llama-server externally as a background service, and pointed atomic-agent's config.localModels.mode at "external" instead of "managed".

Happy to test a fix against this environment if useful.